GitLab

In fact, tkey_test is currently not using -Wl,--wrap at all, despite #ifdef LD_WRAP and the related configure.ac checks. Here is the culprit:

tkey_test@EXEEXT@: tkey_test.@O@ ${ISCDEPLIBS} ${DNSDEPLIBS}
	if test "${LD_WRAP_TESTS}" = true; then WRAP="${WRAP_OPTIONS}"; fi; \
	${LIBTOOL_MODE_LINK} ${PURIFY} ${CC} ${CFLAGS} \
			${LDFLAGS} ${WRAP} -o $@ tkey_test.@O@ \
			${DNSLIBS} ${ISCLIBS} ${LIBS}

Note that ${WRAP} in the libtool invocation only contains a single dollar sign, which means make will try evaluating a make variable (not an environment variable) called WRAP, which is not defined (and thus evaluates to an empty string).

If lib/dns/tests/Makefile.in is changed to use $${WRAP} instead, tkey_test is still compiled correctly, but the -Wl,--wrap part of the linker invocation is essentially a noop as the actual C code in lib/dns/tests/tkey_test.c does not depend on -Wl,--wrap to work.

The important point here is that -Wl,--wrap does not work with calls made from inside shared libraries (it is not like LD_PRELOAD). Thus, since tkey_test calls e.g. dns_tkeyctx_create(), which resides inside libdns, and tkey_test does not call isc__mem_get() directly, and libdns is not built with -Wl,--wrap, dns_tkeyctx_create() still calls the "real" isc__mem_get() from libisc.

What makes tkey_test work are its local definitions of isc__mem_*() functions that call __wrap_isc__mem_*() functions directly. However, these local versions of isc__mem_*() do not need -Wl,--wrap support in the linker in order to work properly.

All in all, I think we need to reconsider our strategy here, i.e. I think:

• we do not need -Wl,--wrap support in the linker to achieve what we want in tkey_test,
• in the typical case, we cannot use -Wl,--wrap with unit tests linked with BIND shared libraries to achieve what we want.

@ondrej, thoughts?